How Visual Experience Shapes Face Processing: Divergent Representational Strategies Emerge from Specialized and Diverse Visual Diets in Artificial Neural Networks

Face perception in humans is supported by specialized neural mechanisms, yet the developmental origins of these mechanisms remain debated. We trained two artificial neural networks (ANNs) on identical orientation-classification tasks but with distinct visual diets: one exposed exclusively to adult faces (Face-ANN) and the other to diverse object categories (Object-ANN). Face-ANN acquired human-like face-processing signatures, including robust performance on Mooney faces, infant faces, and minimal three-dot patterns, and a reliance on low-frequency global structure. However, it failed to recognize face-like objects or most animal faces (except for monkeys), indicating limited overgeneralization. Object-ANN showed the opposite profile, excelling at face-like objects and animal faces but failing on abstract faces. These dissociations demonstrate that domain-specific visual experience alone can give rise to core properties of human face perception, while broad visual experience supports flexible generalization and face pareidolia. Our findings highlight how distinct visual diets shape representational strategies and offer a computational framework for probing the evolutionary origins of face-selective systems.